The lack of data transparency in Large Language Models (LLMs) has highlighted the importance of Membership Inference Attack (MIA), which differentiates trained (member) and untrained (non-member) data. Though it shows success in previous studies, recent research reported a near-random performance in different settings, highlighting a significant performance inconsistency. We assume that a single setting doesn’t represent the distribution of the vast corpora, causing members and non-members with different distributions to be sampled and causing inconsistency. In this study, instead of a single setting, we statistically revisit MIA methods from various settings with thousands of experiments for each MIA method, along with study in text feature, embedding, threshold decision, and decoding dynamics of members and non-members. We found that (1) MIA performance improves with model size and varies with domains, while most methods do not statistically outperform baselines, (2) Though MIA performance is generally low, a notable amount of differentiable member and non-member outliers exists and vary across MIA methods, (3) Deciding a threshold to separate members and non-members is an overlooked challenge, (4) Text dissimilarity and long text benefit MIA performance, (5) Differentiable or not is reflected in the LLM embedding, (6) Member and non-members show different decoding dynamics.

With the development of large language models (LLMs), social biases in these LLMs have become a pressing issue.Although there are various benchmarks for social biases across languages, the extent to which Japanese LLMs exhibit social biases has not been fully investigated.In this study, we construct the Japanese Bias Benchmark dataset for Question Answering (JBBQ) based on the English bias benchmark BBQ, with analysis of social biases in Japanese LLMs.The results show that while current open Japanese LLMs with more parameters show improved accuracies on JBBQ, their bias scores increase.In addition, prompts with a warning about social biases and chain-of-thought prompting reduce the effect of biases in model outputs, but there is room for improvement in extracting the correct evidence from contexts in Japanese. Our dataset is available at https://github.com/ynklab/JBBQ_data.

An growing number of studies have examined the social bias of rapidly developed large language models (LLMs). Although most of these studies have focused on bias occurring in a single social attribute, research in social science has shown that social bias often occurs in the form of intersectionality—the constitutive and contextualized perspective on bias aroused by social attributes. In this study, we construct the Japanese benchmark inter-JBBQ, designed to evaluate the intersectional bias in LLMs on the question-answering setting. Using inter-JBBQ to analyze GPT-4o and Swallow, we find that biased output varies according to its contexts even with the equal combination of social attributes.

Large Language Models (LLMs) can generate the same sequences contained in the pre-train corpora, known as memorization.Previous research studied it at a macro level, leaving micro yet important questions under-explored, e.g., what makes sentences memorized, the dynamics when generating memorized sequence, its connection to unmemorized sequence, and its predictability.We answer the above questions by analyzing the relationship of memorization with outputs from LLM, namely, embeddings, probability distributions, and generated tokens.A memorization score is calculated as the overlap between generated tokens and actual continuations when the LLM is prompted with a context sequence from the pre-train corpora.Our findings reveal:(1) The inter-correlation between memorized/unmemorized sentences, model size, continuation size, and context size, as well as the transition dynamics between sentences of different memorization scores,(2) A sudden drop and increase in the frequency of input tokens when generating memorized/unmemorized sequences (boundary effect),(3) Cluster of sentences with different memorization scores in the embedding space,(4) An inverse boundary effect in the entropy of probability distributions for generated memorized/unmemorized sequences,(5) The predictability of memorization is related to model size and continuation length. In addition, we show a Transformer model trained by the hidden states of LLM can predict unmemorized tokens.

Storytelling is multi-modal in the real world. When one tells a story, one may use all of the visualizations and sounds along with the story itself. However, prior studies on storytelling datasets and tasks have paid little attention to sound even though sound also conveys meaningful semantics of the story. Therefore, we propose to extend story understanding and telling areas by establishing a new component called background sound which is story context-based audio without any linguistic information. For this purpose, we introduce a new dataset, called Sound of Story (SoS), which has paired image and text sequences with corresponding sound or background music for a story. To the best of our knowledge, this is the largest well-curated dataset for storytelling with sound. Our SoS dataset consists of 27,354 stories with 19.6 images per story and 984 hours of speech-decoupled audio such as background music and other sounds. As benchmark tasks for storytelling with sound and the dataset, we propose retrieval tasks between modalities, and audio generation tasks from image-text sequences, introducing strong baselines for them. We believe the proposed dataset and tasks may shed light on the multi-modal understanding of storytelling in terms of sound.

In this paper, we evaluate the progress of our field toward solving simple factoid questions over a knowledge base, a practically important problem in natural language interface to database. As in other natural language understanding tasks, a common practice for this task is to train and evaluate a model on a single dataset, and recent studies suggest that SimpleQuestions, the most popular and largest dataset, is nearly solved under this setting. However, this common setting does not evaluate the robustness of the systems outside of the distribution of the used training data. We rigorously evaluate such robustness of existing systems using different datasets. Our analysis, including shifting of training and test datasets and training on a union of the datasets, suggests that our progress in solving SimpleQuestions dataset does not indicate the success of more general simple question answering. We discuss a possible future direction toward this goal.

Many researchers have tried to predict the accuracies of extrinsic evaluation by using intrinsic evaluation to evaluate word embedding. The relationship between intrinsic and extrinsic evaluation, however, has only been studied with simple correlation analysis, which has difficulty capturing complex cause-effect relationships and integrating external factors such as the hyperparameters of word embedding. To tackle this problem, we employ partial least squares path modeling (PLS-PM), a method of structural equation modeling developed for causal analysis. We propose a causal diagram consisting of the evaluation results on the BATS, VecEval, and SentEval datasets, with a causal hypothesis that linguistic knowledge encoded in word embedding contributes to solving downstream tasks. Our PLS-PM models are estimated with 600 word embeddings, and we prove the existence of causal relations between linguistic knowledge evaluated on BATS and the accuracies of downstream tasks evaluated on VecEval and SentEval in our PLS-PM models. Moreover, we show that the PLS-PM models are useful for analyzing the effect of hyperparameters, including the training algorithm, corpus, dimension, and context window, and for validating the effectiveness of intrinsic evaluation.